cfierro/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k
The cfierro/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k model is an 8 billion parameter Llama-3.1-Instruct variant fine-tuned by cfierro. It specializes in Othello gameplay, trained on 500,000 Othello games across multiple move notation formats. This model is designed to accurately predict legal Othello moves and understand game states, offering a specialized tool for Othello AI research and development.
Loading preview...
Model Overview
This model, cfierro/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k, is an 8 billion parameter Llama-3.1-Instruct model that has undergone full fine-tuning (FFT) specifically for Othello gameplay. It was trained by cfierro using the Axolotl framework.
Key Capabilities
- Othello Game Understanding: The model is fine-tuned on 500,000 Othello games from the
cfierro/othello-llama3-multi-formatdataset, enabling it to process and understand Othello game states. - Multi-Format Move Notation: It supports and was trained on seven different move notation formats, allowing for flexible input and output representations of Othello moves.
- Instruction Following: Alongside Othello data, the model was trained with a 2:1 game-to-chat replay ratio using a subset of
cfierro/tulu3-sft-replay-othello-500kto maintain general instruction-following capabilities. - Legal Move Prediction: The primary objective of the training was to improve the model's ability to predict legal Othello moves, addressing the complexity of Othello's global bracketing rules.
Training Details
The model was trained for one epoch over 500,000 Othello games, which translates to approximately 8,514 steps. The training utilized a learning rate of 1e-5 and employed DeepSpeed ZeRO Stage 3 for efficient multi-GPU training. Validation was performed on a held-out set of 500 Othello games, with the best model selected based on validation loss, which reached 0.7790.